NVH signal processing method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides an NVH signal processing method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining an NVH signal; performing time-frequency transformation on the NVH signal to obtain a time-frequency diagram; and processing the time-frequency graph by using a target ridge extraction model to obtain an order ridge, the target ridge extraction model being obtained by training a pre-constructed soft strategy reinforcement learning model based on an offline data set and a mixed reward function, and the mixed reward function being obtained by constructing based on energy density convolution and visual distance deviation. According to the method, the technical problem that the precision of an existing ridge line extraction method is insufficient under noise interference and cross order interference is solved.

Description

Technical Field

[0001] This application relates to the field of NVH technology, and in particular to an NVH signal processing method, apparatus, electronic device, and storage medium. Background Technology

[0002] In the field of automotive NVH (noise, vibration, and harshness) signal analysis, for the identification and diagnosis of non-stationary signals from rotating machinery, such as vibration and noise sources in key components like engines and transmissions, traditional order tracking and ridge extraction techniques mainly rely on methods like Short-Time Fourier Transform (STFT), wavelet transform, and Volk-Kalman filters. These methods extract ridges by searching for time-frequency energy peaks. However, in practical engineering applications, signals are often affected by complex background noise, order crossovers, and instantaneous impacts, leading to significant limitations in the accuracy of existing techniques. Summary of the Invention

[0003] This application provides an NVH signal processing method, apparatus, electronic device, and storage medium, aiming to improve the technical problem of insufficient accuracy of existing ridge extraction methods under noise interference and cross-order interference.

[0004] According to one aspect of the embodiments of this application, an NVH signal processing method is provided, comprising: acquiring an NVH signal; performing time-frequency transformation on the NVH signal to obtain a time-frequency map; and processing the time-frequency map using a target ridge extraction model to obtain order ridges, wherein the target ridge extraction model is obtained by training a pre-constructed soft policy reinforcement learning model based on an offline dataset and a hybrid reward function, and the hybrid reward function is constructed based on energy density convolution and visual distance bias.

[0005] The NVH signal processing device provided in this application achieves the following technical effects: First, it acquires NVH (noise, vibration, and acoustic roughness) signals, which are the basic data for the diagnosis and analysis of rotating machinery. Next, these signals are converted into time-frequency graphs, a two-dimensional image that intuitively displays the signal energy distribution over time. Finally, using a soft-policy reinforcement learning model, through training and optimization, the time-frequency graph is processed to identify and trace the order ridges of the signal, i.e., the feature paths in the signal related to the rotational speed of the rotating component. The construction of the hybrid reward function considers energy density convolution (i.e., the distribution of signal intensity) and visual distance bias (i.e., the continuity and stability of the ridge path) to guide the model's learning. The above embodiment, by introducing a soft-policy reinforcement learning model and a hybrid reward function, achieves robust extraction of rotating machinery signal features. The improvements in this application's embodiment address the shortcomings of traditional methods in noise interference and order intersection scenarios, significantly improving signal analysis accuracy, and thus solving the technical problem of insufficient accuracy in existing ridge extraction methods under noise interference and order intersection interference.

[0006] Optionally, the training of the target ridge extraction model includes: defining the state space and action space of the soft policy reinforcement learning model to obtain the first defined parameters; defining a hybrid reward function based on energy density convolution and visual distance bias to obtain the second defined parameters; and training the soft policy reinforcement learning model based on the first defined parameters, the second defined parameters, and the offline dataset to obtain the target ridge extraction model.

[0007] Optionally, a target ridge extraction model is obtained by training a soft policy reinforcement learning model based on the first defined parameters, the second defined parameters, and the offline dataset, including: defining the multilayer perceptron and activation function in the soft policy reinforcement learning model; and training the soft policy reinforcement learning model to obtain the target ridge extraction model based on the multilayer perceptron, activation function, first defined parameters, second defined parameters, and the offline dataset.

[0008] Optionally, the state space and action space of the soft policy reinforcement learning model are defined to obtain the first defined parameters, including: defining the normalized energy value of the local window of the time-frequency map and the visual distance of the historical ridge path as the state space; defining the lateral and longitudinal offsets of the ridges on the time-frequency map as the action space; and determining the first defined parameters based on the state space and action space.

[0009] Optionally, training a pre-built soft-policy reinforcement learning model includes: training the pre-built soft-policy reinforcement learning model using an entropy regularization strategy.

[0010] Optionally, the signal processing method further includes: performing NVH diagnosis on the mechanical components corresponding to the order ridge line according to the order ridge line, and obtaining the diagnosis result; and outputting a result prompt signal according to the diagnosis result.

[0011] Optionally, the time-frequency graph is processed using a target ridge extraction model to obtain order ridges, including: processing the time-frequency graph using the target ridge extraction model to obtain an optimal action sequence; determining candidate ridges based on the optimal action sequence; and filtering and eliminating the candidate ridges to obtain order ridges.

[0012] According to another aspect of the embodiments of this application, an NVH signal processing apparatus is also provided, comprising: an acquisition module for acquiring NVH signals; a transformation module for performing time-frequency transformation on the NVH signals to obtain a time-frequency map; and a processing module for processing the time-frequency map using a target ridge extraction model to obtain order ridges, wherein the target ridge extraction model is obtained by training a pre-constructed soft policy reinforcement learning model based on an offline dataset and a hybrid reward function, and the hybrid reward function is constructed based on energy density convolution and visual distance bias.

[0013] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the NVH signal processing method of any of the above.

[0014] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, in which a computer program is stored, the computer program being configured to execute the vehicle fault decision method described above when run on a computer or processor. Attached Figure Description

[0015] Figure 1 This is a flowchart of an NVH signal processing method provided in an embodiment of this application;

[0016] Figure 2 This is a time-frequency diagram example provided in an embodiment of this application;

[0017] Figure 3 This is a schematic diagram of model training provided in an embodiment of this application;

[0018] Figure 4 This is a schematic diagram of the first tracking effect provided in an embodiment of this application;

[0019] Figure 5 This is a schematic diagram of the second tracking effect provided in an embodiment of this application;

[0020] Figure 6 This is a schematic diagram of the third tracking effect provided in an embodiment of this application;

[0021] Figure 7 This is a structural diagram of the NVH signal processing device provided in the embodiments of this application;

[0022] Figure 8 This is a structural diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0023] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] This application provides an NVH signal processing method, comprising: acquiring an NVH signal; performing time-frequency transformation on the NVH signal to obtain a time-frequency map; and processing the time-frequency map using a target ridge extraction model to obtain order ridges. The target ridge extraction model is obtained by training a pre-constructed soft policy reinforcement learning model based on an offline dataset and a hybrid reward function. The hybrid reward function is constructed based on energy density convolution and visual distance bias.

[0026] The NVH signal processing method provided in this application achieves the following technical effects: First, NVH (noise, vibration, and acoustic roughness) signals are acquired, which are the basic data for the diagnosis and analysis of rotating machinery. Next, these signals are converted into time-frequency graphs, a two-dimensional image that intuitively displays the change of signal energy distribution over time. Finally, a soft-policy reinforcement learning model is used to process the time-frequency graph through training and optimization, identifying and tracing the order ridges of the signal, i.e., the feature paths in the signal related to the rotational speed of the rotating component. The construction of the hybrid reward function considers energy density convolution (i.e., the distribution of signal intensity) and visual distance bias (i.e., the continuity and stability of the ridge path) to guide the model's learning. The above embodiment, by introducing a soft-policy reinforcement learning model and a hybrid reward function, achieves robust extraction of rotating machinery signal features. The improvements in this application's embodiment address the shortcomings of traditional methods in noise interference and order intersection scenarios, significantly improving signal analysis accuracy, and thus solving the technical problem of insufficient accuracy of existing ridge extraction methods under noise interference and order intersection interference.

[0027] It should be noted that the following description of the embodiments of the present invention is based on testing the central domain controller of a vehicle.

[0028] This application provides an NVH signal processing method. Please refer to the following embodiments. Figure 1 This includes the following steps:

[0029] Step S100: Obtain NVH signal.

[0030] NVH signals refer to the collective term for noise, vibration, and harshness signals, and are the input data source in this application embodiment, directly derived from the sensor network of the automotive NVH monitoring system. These signals reflect the dynamic information of key components such as the powertrain and suspension system during vehicle operation, and are important basis for analyzing the vehicle's health status and diagnosing potential faults. In this application embodiment, this step ensures the integrity and real-time nature of the basic data for subsequent technical processing.

[0031] Step S200: Perform time-frequency transformation on the NVH signal to obtain a time-frequency diagram;

[0032] Time-frequency transformation (TF-FFT) is a signal processing technique used to convert time-domain signals into time-frequency domain representations, revealing the non-stationary characteristics of the signal. In the embodiments of this application, short-time Fourier transform (STFT) or continuous wavelet transform methods are used to process the acquired NVH signals to generate a time-frequency diagram. Figure 2As shown, the time-frequency plot visualizes the energy distribution of a signal, with the horizontal axis representing time and the vertical axis representing frequency. The value corresponding to each pixel in the image reflects the signal intensity or energy. This transformation allows the order characteristics of rotating machinery (i.e., the frequency components related to rotational speed) to be presented as ridges on the time-frequency plot, providing a clear visual representation for the next step of ridge identification.

[0033] Step S300: The time-frequency map is processed using the target ridge extraction model to obtain the order ridge. The target ridge extraction model is obtained by training a pre-built soft policy reinforcement learning model based on an offline dataset and a hybrid reward function. The hybrid reward function is constructed based on energy density convolution and visual distance bias.

[0034] The target ridge extraction model is based on a soft policy reinforcement learning (SAC) model, combined with a hybrid reward function designed specifically for this task, and trained through deep learning. The SAC model is an advanced reinforcement learning algorithm that can achieve policy optimization in a continuous action space. Through adaptive learning, it finds the optimal action in a given state, balancing exploration and exploitation, and is suitable for handling decision-making problems in complex environments.

[0035] The construction of the hybrid reward function is a unique innovation of this scheme, based on two key concepts: energy density convolution and visual distance bias. Energy density convolution performs convolution operations on the energy distribution of local regions in the time-frequency map, which can focus on areas with high signal intensity and guide the model to learn more accurate ridge paths. Visual distance bias is used to evaluate the continuity and stability of ridges, ensuring that the ridge extraction process will not cause path breaks due to local noise or signal abrupt changes, thereby improving the robustness of order ridge extraction.

[0036] Trained on an offline dataset, the target ridge extraction model learns the optimal strategy for searching and tracking order ridges in time-frequency maps. This process not only considers the time-frequency energy distribution of the signal but also optimizes the visual continuity and path stability of the ridges through a hybrid reward function. After obtaining the trained model, embodiments of this application can process real-time or stored NVH signal time-frequency maps, efficiently and accurately identifying order ridges of rotating machinery, providing crucial information for machinery health status assessment and fault diagnosis.

[0037] The NVH signal processing device provided in this application achieves the following technical effects: First, it acquires NVH (noise, vibration, and acoustic roughness) signals, which are the basic data for the diagnosis and analysis of rotating machinery. Next, these signals are converted into time-frequency graphs, a two-dimensional image that intuitively displays the signal energy distribution over time. Finally, using a soft-policy reinforcement learning model, through training and optimization, the time-frequency graph is processed to identify and trace the order ridges of the signal, i.e., the feature paths in the signal related to the rotational speed of the rotating component. The construction of the hybrid reward function considers energy density convolution (i.e., the distribution of signal intensity) and visual distance bias (i.e., the continuity and stability of the ridge path) to guide the model's learning. The above embodiment, by introducing a soft-policy reinforcement learning model and a hybrid reward function, achieves robust extraction of rotating machinery signal features. The improvements in this application's embodiment address the shortcomings of traditional methods in noise interference and order intersection scenarios, significantly improving signal analysis accuracy, and thus solving the technical problem of insufficient accuracy in existing ridge extraction methods under noise interference and order intersection interference.

[0038] Optionally, the training of the target ridge extraction model includes: defining the state space and action space of the soft policy reinforcement learning model to obtain the first defined parameters; defining a hybrid reward function based on energy density convolution and visual distance bias to obtain the second defined parameters; and training the soft policy reinforcement learning model based on the first defined parameters, the second defined parameters, and the offline dataset to obtain the target ridge extraction model.

[0039] In this embodiment, defining the state space and action space of the soft policy reinforcement learning model (SAC) is a prerequisite for constructing the target ridge extraction model. The state space refers to the model's input during the decision-making process. Specifically, in this embodiment, the state space is defined as local window features on the time-frequency graph, including the energy value of the time-frequency graph and the visual distance deviation between the model's current ridge path and historical ridge paths. The originality of this definition lies in introducing the visual distance deviation as part of the state. This not only captures the signal intensity information but also considers the continuity and stability of the ridge path, providing the model with direct feedback on path continuity and significantly improving the model's adaptability to complex signal sequences. The action space is the set of decisions that the model can execute in each state. In this embodiment, the action space is defined as continuous actions controlling the ridge's lateral and longitudinal movement on the time-frequency graph. This allows the model to flexibly explore different regions in the time-frequency graph and find the optimal ridge path, demonstrating the flexibility and accuracy of this embodiment in processing non-stationary signals.

[0040] The hybrid reward function, based on energy density convolution and visual distance bias, aims to guide the model to learn the optimal ridge tracking strategy during training. Energy density convolution quantifies the energy distribution of the signal in the time-frequency plane by convolving the energy distribution of the time-frequency map. This helps the model identify energy peaks in the signal, which are key features of rotating machinery signals. Visual distance bias measures the deviation between the model's predicted ridge path and the actual signal features, ensuring the continuity of the ridge path and its consistency with the signal features. The introduction of the hybrid reward function enables this embodiment to achieve more efficient and accurate ridge tracking under noisy interference and complex operating conditions by balancing local feature capture and global path planning. This solves the training efficiency problem of traditional methods in sparse reward scenarios and improves the accuracy and stability of ridge extraction.

[0041] Based on the aforementioned first and second defined parameters, and combined with an offline dataset containing noise, order crossover, and complex operating conditions, this embodiment trains a soft-policy reinforcement learning model. This process involves interaction between the model and the environment, continuously trial and error, and updating model parameters to maximize the output of the mixed reward function. During training, the policy learned by the model will be able to automatically search for and track the order ridges of rotating machinery signals in the time-frequency graph, maintaining the continuity and accuracy of ridge tracking even under poor signal quality conditions. This training process enhances the application potential of this embodiment in the field of complex signal analysis, providing an advanced automated tool for fault diagnosis and performance evaluation of rotating machinery.

[0042] This application embodiment achieves accurate and robust extraction of order ridges in rotating machinery signals based on a soft-policy reinforcement learning model through the above steps. Specifically, the definition of the state space considers the continuity of signal strength and ridge path, the setting of the action space allows the model to flexibly search in the time-frequency graph, and the introduction of the hybrid reward function ensures that the model can automatically identify and track the key features of rotating machinery signals, i.e., order ridges, through end-to-end learning under noise interference and complex operating conditions.

[0043] Optionally, a target ridge extraction model is obtained by training a soft policy reinforcement learning model based on the first defined parameters, the second defined parameters, and the offline dataset, including: defining the multilayer perceptron and activation function in the soft policy reinforcement learning model; and training the soft policy reinforcement learning model to obtain the target ridge extraction model based on the multilayer perceptron, activation function, first defined parameters, second defined parameters, and the offline dataset.

[0044] In this embodiment, the Multilayer Perceptron (MLP) and activation functions are the core components of the Soft Policy Reinforcement Learning (SAC) model. The MLP is a feedforward neural network consisting of an input layer, one or more hidden layers, and an output layer. Each layer is connected by weights and is used to learn complex mappings from input to output. During the training of the SAC model, the MLP handles the inputs in the state space, namely the normalized energy value of the local window in the time-domain image and the visual distance bias. Through a combination of multilayer linear transformations and nonlinear activation functions, the model can learn the nonlinear relationship between states and actions, thereby making better decisions in the action space.

[0045] Activation functions are used in neural networks to introduce nonlinear characteristics; common examples include ReLU, LeakyReLU, Sigmoid, and Tanh. In this embodiment, LeakyReLU is chosen as the activation function. Its main advantage is that it prevents gradient vanishing, encourages the model to explore more extensively in the early stages of training, avoids the gradient vanishing problem in neural network training, and ensures the learning efficiency and performance of the model when processing complex signals. By reasonably selecting and configuring activation functions, the model can better capture the nonlinear characteristics of rotating machinery signals, thereby accurately predicting and adjusting the ridge position in scenarios with noise interference and cross-order conditions.

[0046] In this embodiment, based on the aforementioned components and parameters, a soft-policy reinforcement learning model is trained using deep learning. The aim is to obtain a target ridge extraction model capable of accurately processing rotating machinery signals and automatically extracting order ridges. During training, the model performs end-to-end learning based on sample signals from an offline dataset. Through continuous policy optimization and action selection, it aims to maximize the score of the hybrid reward function, i.e., simultaneously considering energy density convolution and visual distance bias. The multilayer perceptron and LeakyReLU activation function work together to enable the model to learn effective feature representations and decision-making strategies from complex signals. This training process ensures that the model can stably and accurately track the order ridges of rotating machinery in real-world application scenarios, even when facing high noise and cross-order interference, thus improving the efficiency and robustness of signal analysis.

[0047] By defining a multilayer perceptron and a special activation function, LeakyReLU, in the soft-policy reinforcement learning model, this application introduces a novel mechanism in reinforcement learning model training. This mechanism can effectively handle the nonlinear characteristics of rotating machinery signals while maintaining adaptability to sparse reward scenarios. Combining the definitions of state space (first defined parameters) and action space (action adjustment), and a hybrid reward function based on energy density convolution and visual distance bias (second defined parameters), the model, after training on an offline dataset, can autonomously learn the order features of rotating machinery signals and dynamically search for the optimal ridge path in the time-frequency plot. This technical process not only significantly improves the accuracy and efficiency of rotating machinery signal analysis and reduces reliance on traditional threshold settings and manual feature engineering, but also enhances the model's generalization ability under complex operating conditions, achieving automation and intelligence in ridge extraction.

[0048] Optionally, the state space and action space of the soft policy reinforcement learning model are defined to obtain the first defined parameters, including: defining the normalized energy value of the local window of the time-frequency map and the visual distance of the historical ridge path as the state space; defining the lateral and longitudinal offsets of the ridges on the time-frequency map as the action space; and determining the first defined parameters based on the state space and action space.

[0049] First, the local window features of the time-frequency graph, including the normalized energy value and the visual distance to the historical ridge path, are defined as the state space of the soft-policy reinforcement learning model. Next, the lateral and vertical offsets of the ridges on the time-frequency graph are set as the action space, thereby defining the range of decisions the model may make at each time point.

[0050] Based on the definitions of state space and action space above, the first defined parameters of the soft policy reinforcement learning model are determined. These parameters include the types of state features and the range of action adjustments that the model needs to consider when making decisions, providing a basic framework for the training and optimization of the model.

[0051] The above steps initialize the soft-policy reinforcement learning model, ensuring that the model possesses self-learning and optimization capabilities in the field of rotating machinery signal analysis, especially when facing complex NVH signals. The model determines the optimal offset action of the ridge line on the time-frequency graph by observing input features in the state space, such as the energy distribution and continuity of historical ridge lines in the time-frequency graph, thereby accurately tracking the order ridge lines of the rotating machinery on the time-frequency graph.

[0052] By clearly defining the state space and action space of the soft-policy reinforcement learning model, this embodiment achieves robust tracking and intelligent extraction of order ridges in rotating machinery signals. The definition of the state space considers the continuity of signal strength and ridge path, ensuring the model can effectively identify key features in the signal. The design of the action space allows the model to flexibly adjust the ridge position to adapt to dynamic changes in the signal. The determination of the first defined parameters provides a clear guide for model training, enabling the model to autonomously optimize the ridge tracking process by learning and adjusting the policy in complex signal environments.

[0053] Optionally, training a pre-built soft-policy reinforcement learning model includes: training the pre-built soft-policy reinforcement learning model using an entropy regularization strategy.

[0054] In this embodiment, entropy regularization is integrated into the training process of a soft-policy reinforcement learning model. By incorporating the entropy value of the policy into the optimization objective, the model is encouraged to explore more broadly in the action space, thereby learning more robust and comprehensive decision-making policies when faced with complex rotating mechanical signals. The originality of this strategy lies in its encouragement of the model to find a balance between exploration and exploitation, preventing the model from over-relying on certain action paths early in training and improving the model's learning and adaptability in sparse reward environments.

[0055] In the training process of this application embodiment, an entropy regularization strategy is specifically implemented to improve the learning efficiency and decision robustness of the soft-policy reinforcement learning model. First, the parameters of the soft-policy reinforcement learning model are randomly initialized, including the weights of the policy network and the value network, to prepare the foundation for subsequent training. An offline dataset containing noise, order crossover, and complex conditions is used to construct a training environment, ensuring that the model can learn in scenarios simulating real-world signal analysis tasks. In each iteration of training, the model selects actions (i.e., lateral and longitudinal offsets of the ridges) based on the current state (i.e., local window features of the i.e., the visual distance of the historical ridge path) and updates the model parameters based on the feedback of the actions and the hybrid reward function (based on energy density convolution and visual distance bias). The entropy regularization strategy plays a crucial role in this process. By adding a regularization term for the policy network entropy value to the model's loss function, it encourages the model to explore more possibilities in the action space, avoids premature convergence, ensures that the model can learn a rich action distribution, and improves adaptability and robustness in complex signal environments.

[0056] By employing an entropy regularization strategy to train a pre-built soft-policy reinforcement learning model, the embodiments of this application significantly enhance the accuracy and stability of ridge extraction in rotating machinery signal analysis. The introduction of the entropy regularization strategy ensures the exploratory nature of the model in the action space, avoiding premature convergence to local optima. Especially in scenarios where the signal is subject to noise interference and cross-order conditions exist, the model can more comprehensively learn and understand signal features, dynamically adjust the ridge path, and achieve continuous and accurate ridge tracking.

[0057] Optionally, the signal processing method further includes: performing NVH diagnosis on the mechanical components corresponding to the order ridge line according to the order ridge line, and obtaining the diagnosis result; and outputting a result prompt signal according to the diagnosis result.

[0058] In this application embodiment, NVH (noise, vibration, and harshness) diagnosis of the mechanical components corresponding to the order ridges is a key post-processing technique. Based on the order ridge features precisely extracted from the time-frequency diagram, it performs in-depth analysis of the NVH performance of critical components such as automotive engines and transmissions. This process involves associating the specific frequency components indicated by the order ridges with their corresponding mechanical components. By evaluating the temporal and frequency stability of these frequency components, as well as their relative intensity with other orders, the health status of the components and potential failure modes are determined. For example, abnormal noise or vibration spectra may manifest as discontinuous abrupt changes or abnormal energy spikes on the order ridges. These features can directly point to potential mechanical faults, such as bearing damage or gear wear.

[0059] Outputting a result prompt signal based on the diagnostic results is another important step in this application embodiment. This involves transforming the diagnostic analysis results into an easily understandable and operable information format for technical personnel to reference or for direct system use. The result prompt signal can contain various information, such as the specific identifier of the faulty component, the fault type, severity, and recommended maintenance measures. In practical applications, this signal may be displayed in the form of a graphical interface, such as highlighting abnormal ridge areas on a time-frequency graph, accompanied by a detailed diagnostic report and recommendations; alternatively, it can be a direct alarm signal that triggers the warning mechanism of the automotive electronic system, reminding the driver or maintenance personnel to check the vehicle's status promptly. This technical action aims to simplify complex data analysis results into intuitive operational guidelines, ensuring that this application embodiment not only performs precise signal feature extraction but also further transforms the analysis results into practical maintenance and fault prediction guidelines, improving the efficiency and relevance of automotive NVH management and diagnosis.

[0060] By leveraging the ability to perform NVH (Noise, Vibration, and Harshness) diagnostics on mechanical components corresponding to order ridges, this application provides a solution directly from signal analysis to fault prediction. Through precise tracking and extraction of order ridges in rotating machinery signals, the system can quickly locate abnormal patterns in the signals, associate them with specific mechanical components, and achieve automated NVH performance diagnosis. This technical process eliminates the need for reliance on human experience, reduces subjectivity in the diagnostic process, and ensures high accuracy and timeliness of results. Furthermore, converting diagnostic results into result prompts facilitates technicians' rapid understanding of complex analysis results, or allows for direct response from the onboard system to initiate preventative maintenance or fault repair procedures, significantly improving the intelligence level of vehicle performance monitoring and maintenance.

[0061] Optionally, the time-frequency graph is processed using a target ridge extraction model to obtain order ridges, including: processing the time-frequency graph using the target ridge extraction model to obtain an optimal action sequence; determining candidate ridges based on the optimal action sequence; and filtering and eliminating the candidate ridges to obtain order ridges.

[0062] In this embodiment, a pre-trained target ridge extraction model (based on a SAC reinforcement learning model) is used to process time-frequency image data. This model, trained through deep learning, is capable of intelligently identifying and tracking order ridge features in rotating machinery signals. In specific applications, when the model receives time-frequency image data input, it applies its learned strategy to perform action selection in the action space to maximize the objective defined by the reward function, i.e., finding the optimal ridge path corresponding to the energy peak. By combining the dynamic decision-making mechanism of deep reinforcement learning with complex signal analysis tasks, the model's automatic exploration and decision-making replace traditional threshold setting or manual feature extraction, significantly improving the automation and accuracy of signal analysis.

[0063] The optimal action sequence is a series of actions output by the model during the decision-making process. These actions guide the model to dynamically track the position of ridges on the time-frequency graph. Through end-to-end training of the model, the optimal action sequence ensures that the model makes the best action selection at each time point and frequency interval based on feedback from the state space (i.e., the visual distance deviation between local features of the time-frequency graph and historical ridge paths) and the mixed reward function, thereby depicting the continuous trajectory of the order ridges on the time-frequency graph. This sequence directly reflects the model's understanding and analytical ability of rotating machinery signals and is key to achieving high-precision ridge tracking.

[0064] The optimal action sequence output by the model guides the initial construction of ridges, forming candidate ridges. Candidate ridges are the first basic outlines of the order characteristic paths of the rotating machinery signal on the time-frequency plot based on the optimal action sequence. This step converts continuous actions into continuous coordinate points on the time-frequency plot, initially presenting the frequency characteristics of the rotating machinery signal at different rotational speeds, laying the foundation for subsequent ridge optimization and final extraction.

[0065] The initially generated candidate ridges are filtered to remove noise and discontinuities, resulting in smoother and more accurate ridge lines. This process can be achieved using a sliding window smoothing filter algorithm. By moving a window across the time-frequency graph, averaging or median filtering is applied to the ridge coordinates within the window to eliminate local jitter and noise interference, ensuring the continuity and stability of the ridge path. The importance of this technique lies in its enhanced robustness of ridge extraction. Even with poor signal quality or the presence of overlapping orders, it can output clear and coherent ridge features, providing reliable data support for NVH diagnostics.

[0066] The technical process described in this application directly achieves accurate conversion from time-frequency plot data to order ridges of rotating machinery signals. By processing the time-frequency plot using a target ridge extraction model, the extracted optimal action sequence ensures that the model can adaptively find key signal features even in noisy and cross-order scenarios. By identifying candidate ridges and filtering them out, the final generated order ridges are not only continuous and accurate but also reliably reflect the dynamic characteristics of rotating machinery under complex operating conditions.

[0067] For example, refer to Figure 3 The training and use of the target ridge extraction model are implemented as follows:

[0068] The process begins with building a dataset, laying the foundation for model training. This stage involves collecting and processing NVH signals, including but not limited to vibration and noise signals from key components such as automotive engines and transmissions under different operating conditions and noise levels. Preprocessing techniques, such as synchronous resampling and time-frequency transformation, convert the signals into a time-frequency graph format suitable for model input. Next, sample augmentation is performed, generating more training samples through data augmentation techniques, such as adding random noise, altering signal strength or frequency distribution, to improve the model's generalization ability and ensure stable performance even when facing unknown or complex signals.

[0069] After constructing a sufficient training dataset, the next step is to set up the Training Environment. This environment simulates real-world scenarios for NVH signal analysis, including various operating states and noise conditions of rotating machinery. Through random sampling, different samples are drawn from the dataset in each training iteration to ensure the model is exposed to diverse signal features, promoting the diversity and robustness of the strategies.

[0070] The SAC (Soft Actor-Critic) model employs multi-environment concurrent training, meaning it is trained simultaneously in multiple simulated NVH analysis scenarios. Each environment has its unique signals and disturbances. Within these environments, the model learns to adaptively select actions based on local features of the time-frequency plot and the visual distance to historical ridge paths. For example, it adjusts the lateral or longitudinal offset of the ridges to maximize the score of a hybrid reward function based on energy density convolution and visual distance bias. Multi-environment training not only improves training speed but also enhances the model's adaptability to complex signals.

[0071] Optuna is an advanced hyperparameter optimization tool used to intelligently adjust various parameters of a SAC model during training, such as learning rate, batch size, and discount factor. With Optuna, the model can avoid getting trapped in local optima in the early stages of training, ultimately reaching optimal performance. This process can be visualized as a heatmap, showing the impact of different hyperparameter combinations on model performance. The Optuna strategy gradually identifies the optimal parameter settings over multiple iterations, ensuring both learning efficiency and accurate ridge extraction.

[0072] Once training is complete, the next step is to package and deploy the inference model. This means packaging the trained SAC model (Target Ridge Extraction Model) into a callable service and deploying it to a real-world application environment, such as an automotive NVH monitoring system.

[0073] After the model is deployed, it undergoes multiple rounds of inference, meaning the model continuously processes multiple time-frequency plot data. After each inference, it optimizes itself based on the results, gradually improving the accuracy of ridge extraction. The HIL (Human in the Loop) user-guided mechanism allows professionals to manually label suspected outliers or provide prior knowledge based on the initial ridges output by the model. This feedback is used for subsequent training of the model, further improving its accuracy and applicability.

[0074] Finally, the model performs local optimization and correction of the ridge extraction results. This typically involves filtering and smoothing the initially generated ridges to remove noise and discontinuities, ensuring that the final output ridges are clear and coherent, accurately reflecting the dynamic characteristics of the rotating machinery. This step further enhances the robustness and accuracy of ridge extraction through post-processing techniques, such as sliding window filtering, enabling high-quality analysis results even in cases of poor signal quality or overlapping orders.

[0075] pass Figure 3 This application not only presents the complete process of SAC model construction and deployment, but also emphasizes the importance of dataset construction, multi-environment training, hyperparameter optimization, and user participation, bringing an advanced, robust, and user-friendly automated solution to the field of complex signal analysis. This method significantly improves the accuracy of NVH signal analysis, reduces reliance on human experience, and promotes the intelligent development of rotating machinery health status assessment and fault prediction.

[0076] Reference Figure 4 When using the embodiments of the present invention for single-point unidirectional tracking, the effect is as follows: Figure 4 As shown. (Refer to...) Figure 5 When using the embodiments of the present invention for single-point bidirectional tracking, the effect is as follows: Figure 5 As shown. (Refer to...) Figure 6 When using the embodiments of the present invention for multi-point bidirectional tracking, the effect is as follows: Figure 6 As shown.

[0077] pass Figure 4 , Figure 5 and Figure 6 We can intuitively see the performance of the target ridge extraction model in this application embodiment under different conditions, from single-point unidirectional tracking to complex multi-point bidirectional tracking, as well as the improvement in effect after model optimization. This not only proves the basic function of the model, but also demonstrates its application potential in complex scenarios such as noise interference and cross-order interference.

[0078] This application also provides an NVH signal processing device 70, please refer to... Figure 7 The system includes: an acquisition module 701 for acquiring NVH signals; a transformation module 702 for performing time-frequency transformation on the NVH signals to obtain a time-frequency map; and a processing module 703 for processing the time-frequency map using a target ridge extraction model to obtain order ridges. The target ridge extraction model is obtained by training a pre-built soft policy reinforcement learning model based on an offline dataset and a hybrid reward function. The hybrid reward function is constructed based on energy density convolution and visual distance bias.

[0079] In the above-described device embodiment, firstly, the acquisition module 701 efficiently acquires NVH signals, ensuring the quality and integrity of the signal source. Subsequently, the transformation module 702 converts the acquired non-stationary signals into a time-frequency image. This transformation process preserves the key features of the signal in the time and frequency dimensions, providing a clear visual foundation for subsequent ridge extraction. Most crucially, the processing module 703 utilizes a target ridge extraction model, an innovative tool based on deep reinforcement learning, to intelligently analyze time-frequency images and adaptively track the order ridges in the signal. Even in scenarios with noise interference and complex cross-order conditions, it ensures the accuracy and continuity of the ridges. This model is trained using a hybrid reward function, combining energy density convolution and visual distance bias, significantly enhancing the model's learning efficiency in sparse reward environments, thereby improving the robustness and adaptability of ridge extraction. Through this technological chain, the entire NVH signal analysis process is not only highly automated but also capable of accurately identifying and separating key features in rotating machinery signals, opening up new and efficient pathways for NVH fault diagnosis and vibration noise source identification in the automotive industry, significantly improving the accuracy and real-time performance of signal analysis.

[0080] This application also provides an electronic device 80, please refer to... Figure 8 The system includes a processor 810 and a memory 820. The memory 810 stores computer programs, and the processor 820 executes the programs stored in the memory 810 to implement the NVH signal processing method described in any embodiment of this application. By introducing a soft-policy reinforcement learning model and a hybrid reward function, robust extraction of rotating machinery signal features is achieved. The improvements in the embodiments of this application address the shortcomings of traditional methods in noise interference and order crossing scenarios, significantly improving signal analysis accuracy, and thus solving the technical problem of insufficient accuracy of existing ridge extraction methods under noise interference and order crossing interference.

[0081] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the NVH signal processing method described in any embodiment of this application. By introducing a soft-policy reinforcement learning model and a hybrid reward function, robust extraction of rotating machinery signal features is achieved. The improvements in this application's embodiments address the shortcomings of traditional methods in noise interference and order crossing scenarios, significantly improving signal analysis accuracy, and thus solving the technical problem of insufficient accuracy of existing ridge extraction methods under noise interference and order crossing interference.

[0082] In this application, "multiple" refers to two or more.

[0083] In this application, unless otherwise expressly defined, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0084] The terms “first,” “second,” “third,” “fourth,” etc., in this application (if present) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0085] In this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0086] Unless otherwise specified, all steps in this application may be performed sequentially or randomly. For example, if the method includes steps A and B, it means that the method may include steps A and B performed sequentially, or it may include steps B and A performed sequentially. For example, if the method may also include step C, it means that step C may be added to the method in any order. For example, the method may include steps A, B, and C, or it may include steps A, C, and B, or it may include steps C, A, and B, etc.

[0087] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An NVH signal processing method, characterized in that, include: Acquire NVH signals; The NVH signal is subjected to time-frequency transformation to obtain a time-frequency diagram; The time-frequency map is processed using a target ridge extraction model to obtain order ridges. The target ridge extraction model is obtained by training a pre-built soft policy reinforcement learning model based on an offline dataset and a hybrid reward function. The hybrid reward function is constructed based on energy density convolution and visual distance bias.

2. The NVH signal processing method according to claim 1, characterized in that, The training of the target ridge extraction model includes: Define the state space and action space of the soft policy reinforcement learning model to obtain the first defined parameters; Based on the energy density convolution and the visual distance deviation, the hybrid reward function is defined to obtain the second defined parameters; The target ridge extraction model is obtained by training the soft policy reinforcement learning model based on the first defined parameters, the second defined parameters, and the offline dataset.

3. The NVH signal processing method according to claim 2, characterized in that, The step of training the soft-policy reinforcement learning model to obtain the target ridge extraction model based on the first defined parameters, the second defined parameters, and the offline dataset includes: Define the multilayer perceptron and activation function in the soft policy reinforcement learning model; The target ridge extraction model is obtained by training the soft policy reinforcement learning model based on the multilayer perceptron, the activation function, the first defined parameter, the second defined parameter, and the offline dataset.

4. The NVH signal processing method according to claim 2, characterized in that, The definition of the state space and action space of the soft policy reinforcement learning model yields the first defined parameters, including: The normalized energy value of a local window of the time-frequency graph and the visual distance of the historical ridge path are defined as the state space; The lateral and longitudinal offsets of the ridge line on the time-frequency graph are defined as the action space; The first defined parameter is determined based on the state space and the action space.

5. The NVH signal processing method according to claim 1, characterized in that, The training of the pre-built soft-policy reinforcement learning model includes: An entropy regularization strategy is used to train a pre-built soft-policy reinforcement learning model.

6. The NVH signal processing method according to claim 1, characterized in that, Also includes: Based on the order ridge line, NVH diagnosis is performed on the mechanical component corresponding to the order ridge line to obtain the diagnosis result; Based on the diagnostic results, a result prompt signal is output.

7. The NVH signal processing method according to claim 4, characterized in that, The step of processing the time-frequency graph using the target ridge extraction model to obtain order ridges includes: The time-frequency graph is processed using a target ridge extraction model to obtain the optimal action sequence; Candidate ridge lines are determined based on the optimal action sequence; The candidate ridges are filtered and eliminated to obtain the ridges of the specified order.

8. An NVH signal processing device, characterized in that, include: The acquisition module is used to acquire NVH signals; The transformation module is used to perform time-frequency transformation on the NVH signal to obtain a time-frequency diagram; The processing module is used to process the time-frequency map using a target ridge extraction model to obtain order ridges. The target ridge extraction model is obtained by training a pre-built soft policy reinforcement learning model based on an offline dataset and a hybrid reward function. The hybrid reward function is constructed based on energy density convolution and visual distance bias.

9. An electronic device, characterized in that, Including processor and memory, among which Memory, used to store computer programs; A processor for executing a program stored in memory to implement the method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 7.

Images